That is like saying humans need to sleep, eat, and drink. True, yes. Useful, no.
The more important question is: what does “high quality” actually mean in venture?
For us, it means the potential to become uniquely world class.
A company that can become the clear winner in an important category. A company with real moats. A company that can build something enormous.
This is why we spend so much time trying to understand what is truly unique about a company. Not what is interesting. Not what demos well. Not what sounds differentiated in a pitch deck. Not what the ARR is today. What is actually hard to replicate? What gets stronger over time? What creates a widening gap versus everyone else? And the only way to know is to spend time with these deep tech founders and really understand how the technology, the product, and the company work.
I have written a long blog post on this topic on Two Small Fish’s website. Here is the link.
We are super excited to share that Two Small Fish led YScope’s US$3.9 million financing, with Snow Angels (the Snowflake alumni investment syndicate), Next Wave NYC, UTEST, and other successful founders participating.
YScope was cofounded by University of Toronto Professor Ding Yuan, who is also CEO, Professor Michael Stumm, Dr. Kirk Rodrigues, Dr. David Lion, Yu (Jack) Luo, and Beverly Xu (Guangji Xu). It is a deeply impressive team building open-source logging infrastructure for the AI era, combining deep systems research with real-world production traction.
Its core technology, CLP (Compressed Log Processor), makes log storage, search, and analytics dramatically more efficient for both humans and AI, across cloud and edge environments.
We believe this is a massive opportunity. As the cost of intelligence collapses, AI agents, robots, autonomous vehicles, and other intelligent systems will generate orders of magnitude more machine-generated events. A robotic finger moves. A self-driving car makes a slight turn. An AI agent retries a task. Each action creates an event, and the infrastructure layer that can handle that explosion efficiently will matter enormously.
YScope is also a strong mutual fit for TSF. We invest in the next frontier of computing and its applications, and we know firsthand how painful logging becomes at scale. I have spent enough time with logs that I will never get back. At Wattpad, logging every tap, swipe, and click could easily add up to billions of events a day. That is why YScope’s traction is so compelling, from powering Uber’s production logging platform to operating across more than 1.5 million connected electric vehicles and being used by Fortune 500 organizations.
Congrats to Ding, Michael, Kirk, David, Jack, Beverly, and the entire YScope team. Full blog post here.
Today is Pi Day, and it feels like a good excuse to reflect on an old friend.
Most people say goodbye to our friend π after school. I’ve been lucky enough to stay in touch. The relationship has evolved over the years, from a childhood friendship in math class to something that followed me into engineering school and later into my work. It is a good reminder that the academic foundations we build early do not stay behind. They continue to shape how we see the world and how we build what comes next.
At Two Small Fish, the next frontier of computing is our investment thesis. We see it taking shape across five areas: Vertical AI Platforms, Physical AI, AI Infrastructure, Advanced Computing Hardware, and Smart Energy. For Pi Day, I thought it would be fun to pick one equation I learned along the way for each of these five areas, and reflect on how it still connects to the technologies shaping this next frontier.
For Vertical AI Platforms, I think of the Gaussian distribution: f(x) = 1/(σ√(2π)) · e^(-(x-μ)^2 / 2σ²), which is foundational in probability and statistics. Even as AI becomes more vertical and more embedded in real workflows, it still rests on probability, statistics, and uncertainty. π is there too.
In Physical AI, the equation I think of is ω = 2πf, which defines angular frequency. I studied control systems and, one summer during my junior year at university, wrote software to control a robotic arm. That was an early lesson that once software meets motion, π becomes part of how the physical world is described.
In AI Infrastructure, I think of the Fourier transform: X(f) = ∫ x(t)e^(-j2πft) dt. I studied signal processing, my bachelor’s thesis was in image processing, and my master’s thesis was on noisy CDMA wireless networks. That math shaped how I thought about signals, images, noise, and communication then, and Fourier shows up in modern LLMs now.
In Advanced Computing Hardware, my equation is ℏ = h/2π. I studied optics in communications, which included a lot of quantum mechanics, so Planck’s constant was part of the vocabulary of the field. What stayed with me is that π shows up at the quantum level as part of the structure, not just the math.
In Smart Energy, the equation I would use is Xₗ = 2πfL, which calculates inductive reactance. It is a simple reminder that in AC systems, frequency directly shapes behaviour. As energy systems become smarter and more dynamic, π remains embedded in the physics underneath.
That may be why Pi Day still resonates with me. π is one of those rare constants that keeps reappearing across disciplines, from robotics to quantum mechanics, from signal processing to energy systems, and now across the next frontier of computing.
P.S. I also realized I missed mentioning my other friend Euler back on February 7. Next time!
OpenClaw, an AI agent that can operate a computer on your behalf, has taken the world by storm. Unless you have been living under a rock, you have probably either tried it already or at least wanted to find out what all the buzz is about.
Many, however, have failed to get past installation because it is so difficult. There is a reason why thousands of people lined up for help just to get OpenClaw installed on their machines. More importantly, using it without proper safeguards can create a real security risk.
From my perspective, three issues stand out in OpenClaw’s current form.
First, it is difficult to install, even for technical users. That matters more than many builders realize. A product does not become broadly useful simply because it is powerful. It becomes useful when people can actually get it running without friction or handholding.
Second, it can create a real security risk if not used properly. Tools that operate at the machine level can be compelling, but they also introduce a very different level of responsibility. Most users do not want to expose their full machine environment just to perform a simple task.
Third, it can become expensive quickly. Token bills can become material before users even realize it. A tool may look impressive in a demo, but if the economics do not work, adoption will eventually stall. In AI, performance matters, but efficiency matters just as much.
This is why, after looking at many options, I chose to use Crate from our portfolio company, Gensee, myself, and I believe it is by far the best way to try OpenClaw.
It addresses all three issues directly: one-click install in 60 seconds, a secure sandbox that only accesses what you explicitly allow, and deep expertise from Dr. Shengqi Zhu and award-winning operating systems expert Professor Yiying Zhang, whose work on agentic optimization and efficiency is exactly what makes this possible. That expertise is also why they have been able to make Crate completely free to use.
In other words, it makes OpenClaw easy, safe, and completely free.
There is also a bonus. Crate comes with Gensee’s proprietary AI search engine built in. That search engine ranked #1 on Source Bench for finding the highest-quality web sources.
Another bonus is that Crate comes pre-installed with a set of common, useful skills vetted by the Gensee team for safety, while still allowing users to install additional skills themselves. That makes it both easier to get started and more flexible over time.
A final bonus is flexible control. Users can create multiple instances, pause and resume them, take snapshots, and roll back at any time. That means full control without the usual complexity.
So Gensee Crate is not just an easier and safer way to use OpenClaw. It is also a better one, and that points to where this market is going. The first wave of a technology shows what is possible; the next wave makes it practical for mainstream users. AI agents are now entering that phase. To become part of everyday workflows, they need to be easy to use, safe by design, and efficient enough to be economically viable. That is where adoption happens.
And that is why Gensee Crate is the best way to try out OpenClaw and why it is worth paying attention to.
If you are curious about OpenClaw, try Gensee Crate here.
At Two Small Fish Ventures, we invest in the next frontier of computing and its applications. Supporting that thesis is our focus on research grounded innovation, which means we spend a lot of time with people who are building from first principles and turning technical breakthroughs into real companies. Not surprisingly, many of those people are world-class women researchers, scientists, and engineers. We have been fortunate to back a good number of them, and we are better for it.
This shows up in our portfolio, but it also shows up in our own team. Roughly half of our team is female. Our investment team is also roughly half female, with Eva and Mikayla bringing perspectives that genuinely shape how we think, how we evaluate, and how we support founders.
This is not just something to celebrate. It makes us better. One of the most common pieces of feedback we hear from founders is that we ask very different questions. That is exactly the point. Different perspectives lead to better conversations, smaller blind spots, and stronger judgment. In deep tech, where the path from breakthrough to company is rarely straightforward, that matters.
So today, we celebrate the many women founders, researchers, scientists, and engineers we have backed, and the many more we hope to back in the years ahead.
AI has a massive efficiency problem. It uses too much compute. It costs too much. It uses too much energy. And it is too slow.
Today, it can take a serious cluster of GPUs and a very non-trivial amount of electricity just to answer a simple question like “Can you summarize this document?” or “What should I reply to this email?” The machinery underneath is anything but.
This is why we invested in ByteShape. The company was co-founded by a world-class team out of the University of Toronto: Professor Andreas Moshovos [link]—whose group’s papers have amassed more than 10,000 citations—together with scientists Miloš Nikolić [link], Enrique Torres Sánchez [link], and Ali Hadi Zadeh [link], whose life’s work is making computation more efficient. Both Ali and Miloš were also postgraduate affiliates of the Vector Institute, and Miloš’s PhD research formed the foundation of ByteShape’s core technology—work that earned him recognition as an “ML and Systems Rising Star” by MLCommons last year.
They are building the kind of deep technology that changes the economics of AI deployment, then changes what products become possible.
Quantization, In Plain English
Many techniques underpin what ByteShape does. One of them jumped out: quantization.
Quantization is about using fewer bits to represent the numbers inside a model. Many models are trained with higher precision formats because it helps learning remain accurate. But AI inference often does not need that much precision everywhere. If you can safely represent weights and activations with fewer bits, you shrink memory use and speed up compute, often dramatically, while keeping outputs essentially the same.
ByteShape’s approach, ShapeLearn, does this in a way that feels obvious in hindsight and very hard in practice. ShapeLearn adaptively taps into the AI training process to learn optimal datatypes for parameters and inputs. The result can be 7x faster training and 10x faster inference.
In layman’s terms, the idea is simple and powerful: fewer bits, less work, and smaller models, without sacrificing results. All is being done adaptively.
Then ByteShape takes it one step further. ShapeSqueeze is their lossless compression layer that applies per-value encoding to minimize off-chip data transfers, with up to 40% extra compression.
Put the two together, and you get something that really matters in the real world. ShapeLearn reduces what the model needs to store and compute. ShapeSqueeze reduces what the hardware needs to move around. Less compute and less data movement means faster AI, lower cost, and lower energy.
This is not limited to savings in cloud data centres. It is a step-function improvement in what can run locally, which means a step-function improvement in what products can exist. It opens the door to privacy-sensitive and offline workflows, on-device agents, and embedded intelligence in robots and machines where speed, power and thermals matter.
Why TSF invested
Two Small Fish Ventures is an early-stage deep tech venture capital firm investing globally in the next frontier of computing and its applications. We invest where foundational breakthroughs create the conditions for new category-defining companies, and we back founders at the earliest stages when the technology is ready for commercialization.
ByteShape fits that thesis perfectly. They are building a foundational efficiency layer for AI that can reshape performance and cost across cloud, enterprise, and edge deployments. And because all TSF partners are engineers with deep operating experience, we do not just evaluate the science. We lean into technology through commercialization, with hands-on support informed by having built and scaled companies ourselves.
With ByteShape, the future is models that run faster, use less energy, and fit on far smaller hardware, without sacrificing the quality that makes them worth using.
